File size: 8,297 Bytes
93cd35a | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 | import logging
import os
import re
from typing import Any, Optional, Tuple
import torch
from utils.logging_utils import log_for_0, _process_index
from utils.train_utils import unwrap_model
def _local_path(path: str) -> str:
return os.path.abspath(os.path.expanduser(path))
def upload_output_dir_to_hf(output_dir: str, hf_repo_id: Optional[str], reason: str = "artifacts"):
if not hf_repo_id or _process_index() != 0:
return
folder_path = _local_path(output_dir)
if not os.path.isdir(folder_path):
log_for_0(f"HF upload skipped; output directory does not exist: {folder_path}",
level=logging.WARNING)
return
try:
from huggingface_hub import HfApi
repo_id = hf_repo_id.strip("/")
api = HfApi()
api.create_repo(repo_id, repo_type="model", exist_ok=True)
log_for_0(f"Uploading {reason} to HF: {repo_id}")
api.upload_folder(repo_id=repo_id, folder_path=folder_path, repo_type="model")
log_for_0(f"Uploaded {reason} to HF: {repo_id}")
except Exception as e:
log_for_0(f"Failed to upload {reason} to HF: {e}", level=logging.WARNING)
def _split_hf_path(path: str, min_parts: int) -> Optional[Tuple[str, str]]:
if "://" in path:
return None
if path.startswith(("/", ".", "~")):
return None
if os.path.exists(_local_path(path)):
return None
parts = path.split("/")
if len(parts) < min_parts:
return None
return "/".join(parts[:2]), "/".join(parts[2:])
def save_checkpoint(state, output_dir: str, step: int, hf_repo_id: str = None):
"""Save model checkpoint locally as a single `checkpoint_<step>` file."""
if _process_index() != 0:
return
ckpt_dir = _local_path(output_dir)
os.makedirs(ckpt_dir, exist_ok=True)
inner_model = unwrap_model(state.model)
grad_accum_buffers = {}
if getattr(state, "grad_accum_buffers", None):
for name, param in inner_model.named_parameters():
buf = state.grad_accum_buffers.get(id(param))
if buf is not None:
grad_accum_buffers[name] = buf.detach().cpu()
payload = {
"params": inner_model.state_dict(),
"ema_params1": state.ema_params1,
"opt_state": state.optimizer.state_dict(),
"lr_scheduler": state.lr_scheduler.state_dict() if state.lr_scheduler is not None else None,
"step": int(state.step),
"epoch": int(state.epoch),
"dropout_rng": (state.dropout_generator.get_state()
if state.dropout_generator is not None else None),
"grad_accum_buffers": grad_accum_buffers,
}
out_path = os.path.join(ckpt_dir, f"checkpoint_{step}")
log_for_0(f"Saving checkpoint to {out_path}")
torch.save(payload, out_path)
log_for_0(f"Checkpoint written to {out_path}")
upload_output_dir_to_hf(output_dir, hf_repo_id, reason="checkpoint")
def _checkpoint_step(checkpoint_name: str) -> int:
"""Extract the trailing checkpoint step from a name; -1 if absent."""
match = re.search(r"(\d+)$", checkpoint_name)
return int(match.group(1)) if match else -1
def find_all_checkpoints(ckpt_dir: str, prefix: str = "checkpoint_"):
"""Find local checkpoint paths in a directory, sorted by step ascending."""
ckpt_dir = _local_path(ckpt_dir)
if not os.path.isdir(ckpt_dir):
return []
names = sorted(
[f for f in os.listdir(ckpt_dir) if f.startswith(prefix)],
key=_checkpoint_step,
)
return [os.path.join(ckpt_dir, name) for name in names]
def find_latest_checkpoint(ckpt_dir: str, prefix: str = "checkpoint_"):
"""Return the latest local checkpoint path, or None."""
all_ckpts = find_all_checkpoints(ckpt_dir, prefix)
return all_ckpts[-1] if all_ckpts else None
def _download_hf_checkpoint(checkpoint_path: str) -> Optional[str]:
hf_path = _split_hf_path(checkpoint_path, min_parts=2)
if hf_path is None:
return None
repo_id, sub_path = hf_path
from huggingface_hub import snapshot_download
log_for_0(f"Downloading checkpoint from HF: {repo_id}" + (f"/{sub_path}" if sub_path else ""))
local_dir = snapshot_download(
repo_id=repo_id, repo_type="model",
allow_patterns=[f"{sub_path}/**"] if sub_path else None,
)
return os.path.join(local_dir, sub_path) if sub_path else local_dir
def _restore_checkpoint(checkpoint_path: str) -> Any:
"""Restore a checkpoint from a file or directory (latest inside dir)."""
local = _local_path(checkpoint_path)
resolved = local
if os.path.isdir(local):
latest = find_latest_checkpoint(local)
if latest is not None and os.path.isfile(latest):
resolved = latest
if os.path.isfile(resolved):
return torch.load(resolved, map_location="cpu")
return None
def _validate_checkpoint(ckpt: Any):
if ckpt is None:
raise ValueError("checkpoint restore returned None")
required_keys = ("params", "opt_state", "step", "epoch")
missing_keys = [key for key in required_keys if key not in ckpt]
if missing_keys:
raise ValueError(f"checkpoint restore missing keys: {missing_keys}")
def load_checkpoint(checkpoint_path: str, state) -> Tuple[Any, int]:
"""Load an ELF checkpoint.
Uses an existing local path first; otherwise tries HF and then local fallback.
"""
log_for_0(f"Loading ELF checkpoint from {checkpoint_path}...")
ckpt, loaded_from = None, None
errors = []
local_path = _local_path(checkpoint_path)
if os.path.exists(local_path):
try:
log_for_0(f"Loading local checkpoint from {local_path}...")
ckpt = _restore_checkpoint(local_path)
_validate_checkpoint(ckpt)
loaded_from = "local"
except Exception as e:
errors.append(f"local: {e}")
if ckpt is None:
try:
hf_path = _download_hf_checkpoint(checkpoint_path)
if hf_path:
log_for_0(f"Loading HF checkpoint from {hf_path}...")
ckpt = _restore_checkpoint(hf_path)
_validate_checkpoint(ckpt)
loaded_from = "HF"
except Exception as e:
errors.append(f"HF: {e}")
log_for_0(f"HF checkpoint restore failed ({e}); falling back to local path.")
if ckpt is None:
raise ValueError(
f"Failed to load checkpoint from {checkpoint_path}. Tried: {'; '.join(errors)}"
)
log_for_0(f"Loaded checkpoint keys: {list(ckpt.keys())}")
inner_model = unwrap_model(state.model)
inner_model.load_state_dict(ckpt["params"])
ema_src = ckpt.get("ema_params1", ckpt["params"])
device_map = {n: p.device for n, p in inner_model.named_parameters()}
for n, b in inner_model.named_buffers():
device_map.setdefault(n, b.device)
fallback_device = next(iter(device_map.values()), torch.device("cpu"))
state.ema_params1 = {
n: t.to(device_map.get(n, fallback_device)) for n, t in ema_src.items()
}
state.optimizer.load_state_dict(ckpt["opt_state"])
if state.lr_scheduler is not None and ckpt.get("lr_scheduler") is not None:
state.lr_scheduler.load_state_dict(ckpt["lr_scheduler"])
state.step = int(ckpt["step"])
state.epoch = int(ckpt["epoch"])
if ckpt.get("dropout_rng") is not None and state.dropout_generator is not None:
try:
state.dropout_generator.set_state(ckpt["dropout_rng"])
except Exception:
pass
if ckpt.get("grad_accum_buffers"):
buffers = ckpt["grad_accum_buffers"]
state.grad_accum_buffers = {}
param_ids = []
for name, param in inner_model.named_parameters():
if not param.requires_grad:
continue
param_ids.append(id(param))
saved = buffers.get(name)
state.grad_accum_buffers[id(param)] = (
saved.to(device=param.device, dtype=param.dtype)
if saved is not None else torch.zeros_like(param)
)
state.grad_accum_param_ids = tuple(param_ids)
step = int(ckpt["step"])
log_for_0(f"Loaded {loaded_from} checkpoint from step {step} (epoch {state.epoch})")
return state, step
|